Neural network predictive control method for high-speed maglev train suspension system

Abstract

The application provides a neural network prediction control method for a high-speed maglev train suspension system, which comprises the following steps: 1) obtaining dynamic response in the running process of the maglev train according to the suspension sensor arranged on the upper part of the maglev train; reading the coil current of the electromagnet in the chopper by using the suspension controller on the train; and recording the data as a sample set for constructing a prediction model; 2) constructing a maglev train response prediction model by using the recorded sample for neural network training, so as to realize online accurate prediction of the response of the train at a future time; 3) calculating a feedforward control amount; 4) calculating a feedback control amount; and 5) outputting a control signal after the feedforward and feedback control amounts are superposed by the suspension controller. The application can effectively suppress the fluctuation of the electromagnet of the maglev train under the aerodynamic load, realize accurate prediction control, and make the maglev train run more stably in suspension.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of levitation control of maglev trains, and particularly relates to a neural network prediction control method for a levitation system of a high-speed maglev train. BACKGROUND

[0002] With the development of railway transportation, people have higher requirements for the running speed and comfort of trains. Maglev trains have been widely concerned since their appearance. The electrodynamic maglev train relies on the attraction between electromagnets and tracks to overcome gravity to achieve levitation. In the running process, it does not contact the ground, eliminating the frictional resistance of the wheel-rail system, and has the advantages of fast speed, more comfort and less wear. However, due to the inherent instability of electromagnetic suspension, the electrodynamic maglev train needs real-time control to achieve stable suspension.

[0003] At present, the maglev train still adopts the traditional PID control algorithm, which is a feedback control algorithm based on the direct decoupling of a single electromagnet. When running under complex conditions, it is easy to cause suspension loss of control and other problems. However, the high-speed maglev train has complex operating conditions and will be affected by aerodynamic load, track disturbance, load fluctuation and other factors, which will affect the suspension stability of the train. Especially when the trains meet or pass through the tunnel, they will also be affected by the obvious aerodynamic impact, causing the train to have large suspension gap fluctuation amplitude, suspension instability and other serious problems.

[0004] Although many scholars have proposed some more advanced control strategies, these algorithms are mostly based on the suspension of a small ball or a single electromagnet, and are mostly suitable for linearized systems or simplified decoupled nonlinear systems, without considering the modal characteristics of the overall vehicle structure. However, the single-point suspension system of the maglev train itself has strong nonlinearity, and there are also nonlinear characteristics between each suspension system. Directly simplifying and decoupling the coupled components is not conducive to the overall performance of the system. Therefore, when analyzing the maglev train, the nonlinear characteristics of the vehicle itself and the surrounding environment must be considered. Not only that, but the existing maglev train suspension control strategies have not gone beyond the scope of feedback control. Feedback control mainly feeds back to the consequences, and the movement that has occurred cannot be changed.

[0005] In summary, in order to ensure the safety and comfort of high-speed maglev trains, new and more intelligent control strategies need to be developed. SUMMARY

[0006] The application provides a neural network prediction control method which can be used for a high-speed maglev train suspension system.

[0007] To solve the above technical problems, the application provides a neural network prediction control method which can be used for a high-speed maglev train suspension system.

[0008] 1) According to the suspension sensor arranged on the maglev train, the dynamic response in the running process of the maglev train is obtained; the coil current of the electromagnet in the chopper is read by the suspension controller on the train; and the data are recorded as a sample set for constructing a prediction model;

[0009] 2) The recorded sample is used for neural network training, and a maglev train response prediction model is constructed to realize online accurate prediction of the response at a future time of the train;

[0010] 3) The suspension gap and coil current signals of the electromagnet are read by the suspension controller, the prediction model is used for online prediction of the response at a future time of the train, and the feedforward control amount is calculated according to the prediction;

[0011] 4) The suspension gap and acceleration signals of the electromagnet are read by the suspension controller, and the feedback control amount is calculated according to the signals;

[0012] 5) The suspension controller outputs the control signal after superimposing the feedforward and feedback control amounts.

[0013] The neural network prediction control method which can be used for a high-speed maglev train suspension system, wherein the step 1) comprises the following steps:

[0014] 1.1) determining the data required for constructing the maglev train response prediction model

[0015] The suspension gap and coil current data of the electromagnet of the maglev train are selected as the sample set of the prediction model; that is, the suspension sensor is arranged on each electromagnet of the maglev train to measure the suspension gap of the electromagnet when the train runs, and the coil current data of the electromagnet in the chopper is read by the suspension controller;

[0016] 1.2) collecting the data sample required for constructing the prediction model

[0017] That is, the running conditions of the train are selected, and the suspension gap and coil current time series of the electromagnet of the maglev train under various motion conditions are recorded.

[0018] The neural network prediction control method for the high-speed maglev train suspension system, wherein the step 2) specifically comprises the following steps:

[0019] 2.1) Construct a neural network forward transmission framework;

[0020] 2.2) Construct a neural network error back transmission framework;

[0021] 2.3) Set the basic information of the neural network framework;

[0022] 2.4) Train the neural network constructed in the steps 2.1) and 2.2) by using the data samples obtained in the step 1), to obtain a maglev train dynamic response prediction model.

[0023] The neural network prediction control method for the high-speed maglev train suspension system, wherein the step 2.1) specifically is to set an improved recurrent neural network-long short-term memory neural network to construct a neural network forward transmission prediction framework.

[0024] The neural network prediction control method for the high-speed maglev train suspension system, wherein the step 2.3) specifically comprises the following steps:

[0025] 2.3.1) Select the electromagnetic suspension gap and the coil current at the time t and a period of time before the time t as the input values of the neural network, and select the suspension gap at the time t+1 as the expected output value Y * Train ,

[0026] 2.3.2) Set the neural network parameters; and the neural network parameters comprise the number of hidden layers, the number of hidden layer cells, the output layer error loss function, the back transmission learning rate η and the training stop error LOSS in the neural network end .

[0027] The neural network prediction control method for the high-speed maglev train suspension system, wherein the step 2.4) specifically comprises the following steps:

[0028] 2.4.1) Read the measured maglev train electromagnetic suspension gap and current data in the step 1) as a sample set of the prediction model, and obtain the input values and the expected output values of the neural network from the sample set according to the method in the step 2.3.1);

[0029] 2.4.2) Initialize the neural network, and assign initial random values to the weight and threshold of each neuron;

[0030] 2.4.3) Calculate the input values according to the neural network forward transmission framework constructed in the step 2.1) to obtain the electromagnetic suspension gap prediction value YTrain , Y Train = [y L+1 y L+2 …y n ], then the calculation error LOSS of the output layer is calculated, as shown in the following formula (8):

[0031]

[0032] 2.4.4) comparing the calculation error LOSS calculated in the step 2.4.3) with the training stop error LOSS end : if LOSS < LOSS end , the training is ended, and the trained prediction model f pre () is obtained; if not, the error is back-propagated, and the error of each neuron in each layer is calculated backwardly;

[0033] 2.4.5) calculating the increments Aw, Ab of the weights and thresholds of each neuron by using the error of each neuron calculated in the step 2.4.4), and updating the weights w and thresholds b of each neuron as w = w + Aw, b = b + Ab, and the training is ended once;

[0034] 2.4.6) taking the weights w and thresholds b trained in the step 2.4.5) as the initial values of the neurons in the next round of training, and repeating the calculation process of the steps 2.4.3) to 2.4.5) again, and continuously modifying the weights and thresholds by training until the error LOSS of the output layer is less than the training stop error LOSS end , and finally obtaining the trained prediction model f pre ();

[0035] 2.4.7) constructing a prediction model by using the trained neural network model obtained in the step 2.4.6), and the expression is as follows:

[0036]

[0037] wherein in the above formula (10), z (t) is the predicted value of the suspension gap calculated at the time t; f pre () is the trained neural network prediction model; X L (t) is a measured data sequence of the electromagnet at the time t; L is the number of rows of historical information included in X L (t); x(t) is the real data measured at the time t, wherein z(t) is the measured suspension gap of the electromagnet at the time t, and i(t) is the measured current of the electromagnet at the time t.

[0038] The neural network prediction control method for the suspension system of the high-speed maglev train, wherein the specific steps of the step 3) are as follows:

[0039] 3.1) The levitation controller reads the electromagnetic levitation gap signal and the coil current signal as the input value of the prediction model, and estimates the future response of the levitation system through the prediction model in step 2).

[0040] 3.2) The feedforward control amount is calculated using the predicted value of the levitation system, as shown in the following formula (11):

[0041]

[0042] wherein K ff is the feedforward gain coefficient; z is the predicted levitation gap; z0 is the rated levitation gap, which is 10 mm; the calculation of the feedforward control amount based on the predicted value of the levitation system is not limited to the form shown in formula (11), and can be replaced by other forms as needed.

[0043] The neural network prediction control method for the levitation system of a high-speed maglev train, wherein the specific steps of step 4) are as follows:

[0044] The feedback control amount is calculated by reading the levitation gap and acceleration measured by the sensor through the feedback module of the levitation controller; the proportional (P), integral (I), and differential (D) linear combinations of the deviation between the electromagnetic levitation gap measurement value and the target value are combined into a control amount by using the PID algorithm for feedback control; the electromagnetic acceleration integral term is used to replace the levitation gap differential term, and the expression of the PID algorithm feedback control logic is:

[0045]

[0046] wherein K p is the proportional coefficient; K i is the integral coefficient; K d is the differential coefficient, z(t) is the levitation gap measured by the sensor; a(t) is the electromagnetic acceleration measured by the sensor; z0 is the rated levitation gap;

[0047] The neural network prediction control method for the levitation system of a high-speed maglev train, wherein the specific steps of step 5) are as follows:

[0048] The feedforward control amount in step 3) and the feedback control amount in step 4) are superimposed to obtain the control amount of the prediction control, as shown in the following formula (13):

[0049] u(t+1) = η ff u ff (t+1) + η fb u fb (t+1) (13);

[0050] ​wherein u ff (t) and η ff are feedforward control quantities and their weight coefficients, u fb (t) and η fb are feedback control quantities and their weight coefficients.

[0051] With the technical solution, the application has the following beneficial effects:

[0052] The application can be used for the neural network prediction control method of the high-speed maglev train suspension system, and the prediction model is established based on the dynamic response of the maglev vehicle under the external disturbance such as the aerodynamic load, so that the coupling effect of each part of the maglev vehicle and the influence of the external load are considered, and the application is more in line with the actual situation.

[0053] According to the characteristics of the nonlinear time-varying of the vehicle dynamic response, the long short-term memory neural network is used to construct the prediction model, the accurate prediction of the suspension gap time sequence of the electromagnet of the maglev train can be realized, and a basis is provided for the feedforward control.

[0054] The long short-term memory neural network is used to create the dynamic response prediction model of the maglev train, the online accurate prediction of the vehicle suspension gap is realized, and on this basis, the predicted value is used for the feedforward control of the current of the electromagnet of the vehicle, the control is applied in advance before the suspension system appears a large fluctuation, and the large fluctuation of the vehicle under the impact load can be avoided.

[0055] In order to prevent the prediction error caused by the model mismatch from leading to the invalidation of the feedforward control and the influence of the disturbance compensation, the feedback control and the feedforward control are applied cooperatively, and the stable prediction control of the suspension system of the high-speed maglev train can be realized.

[0056] Compared with the existing control algorithm, the advantages of the feedforward control and the feedback control are combined, the prediction control of the suspension system of the maglev train can be realized, the fluctuation of the suspension gap of the electromagnet of the maglev train under the aerodynamic load can be effectively inhibited, the suspension stability of the high-speed maglev train during operation is improved, and the application has a high advantage compared with the general feedback control strategy. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the specific embodiments of the application or the technical solutions in the prior art, the drawings needed in the following specific embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0058] Figure 1 The flowchart of the neural network prediction control method of the high-speed maglev train suspension system of the application;

[0059] Figure 2 This is a schematic diagram of the LSTM neural network structure in the neural network predictive control method of the high-speed maglev train suspension system of the present invention;

[0060] Figure 3 This is a schematic diagram of the control quantity calculation in the neural network predictive control method for the high-speed maglev train suspension system of the present invention. Detailed Implementation

[0061] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The present invention will be further explained below with reference to specific embodiments.

[0063] This embodiment provides a neural network predictive control algorithm for a high-speed maglev train levitation system. The algorithm combines the advantages of feedforward and feedback to control the electromagnet current, which can effectively improve the levitation stability of the train.

[0064] like Figure 1 As shown, the neural network predictive control algorithm for the high-speed maglev train levitation system of the present invention specifically includes the following steps:

[0065] 1) Based on the suspension sensors deployed on the maglev train, obtain the dynamic response of the maglev train during operation; use the suspension controller on the train to read the coil current of the electromagnet in the chopper; and record these data as a sample set for building the prediction model.

[0066] 1.1) Determine the data required to construct the maglev train response prediction model.

[0067] This invention selects the levitation gap and coil current data of the electromagnets of maglev trains as the sample set for the prediction model; that is, levitation sensors are arranged on each electromagnet of the maglev train to measure the levitation gap of the electromagnets when the train is running, and the coil current data of the electromagnets in the chopper are read by the levitation controller.

[0068] 1.2) Collect data samples required for building the prediction model

[0069] Various operating conditions were selected, including train operation on open tracks, passing other trains, passing through tunnels, straight-line operation, and curve operation. The levitation gap and coil current time series of the maglev train under these motion conditions were recorded.

[0070] 2) Carrying out neural network training, constructing a maglev train response prediction model, and realizing online accurate prediction of train future time response;

[0071] 2.1) Constructing a neural network forward transmission framework; the maglev train dynamic response belongs to a time series, and an improved recurrent neural network-long short-term memory (LSTM) neural network is set to construct a neural network forward transmission prediction framework;

[0072] The neural network structure in the above step 2.1) mainly includes three layers: an input layer, a hidden layer, and a linear output layer; the maglev train dynamic response belongs to a time series, and an improved recurrent neural network-long short-term memory (LSTM) neural network is set to construct a neural network forward transmission prediction framework, and the structure of the LSTM neural network is as shown in Figure 2 .

[0073] The LSTM neural network mainly realizes forward propagation through three gate structures:

[0074] ① The forgetting gate decides how much historical information is forgotten, that is, how much history participates in the calculation at this time:

[0075] f (t) =σ(W fh h (t-1) +b fh +W fx x (t) +b fx ) (1);

[0076] In the above formula (1), f (t) is the output value of the forgetting gate; σ is a sigmoid activation function (the output value is between 0 and 1, and 0 represents complete forgetting); b fh , b fx are the threshold values of the forgetting gate; W fh , W fx are the weights of the forgetting gate; x(t) is the input value at the t-th step, and h (t-1) is the output previous value at the t-th step.

[0077] ② The input gate decides the selection of input information:

[0078] i (t) =σ(W ih h (t-1) +b ih +W ix x (t) +b ix ) (2);

[0079] In the above formula (2), i (t) is the output value of the input gate; W ih , Wix is the weight of the input gate; b ih , b ix is the threshold of the input gate.

[0080] ③Cell state update

[0081] The results of the previous forget gate and input gate will act on the cell state C (t) :

[0082] z (t) = σ(W zh h (t-1) + b zh + W zx x (t) + b zx ) (3).

[0083] C (t) = f (t) ⊙C (t-1) + i (t) ⊙z (t) (4).

[0084] wherein z (t) in the above formula (3) is the cell state update process value; W zh , W zx is the weight of the cell state update; b zh , b zx is the threshold of the cell state.

[0085] ④Output gate, used to determine the value of the next hidden state, which contains the information of the previous input:

[0086] o (t) = σ(W oh h (t-1) + b oh + W ox x (t) + b ox ) (5).

[0087] h (t) = o (t) ⊙ tanh(C (t) ) (6).

[0088] wherein o (t) in the above formula (5) is the output gate process value; W oh , W ox is the weight of the output gate; b oh , b ox is the threshold of the output gate; h (t) in the above formula (6) is the output value of the output gate, that is, the hidden state value;

[0089] (5) Finally, the output value is calculated by the linear layer to obtain the predicted value y pre As shown in the following formula (7):

[0090] y pre = h (t) (W linear x + b linear ) (7);

[0091] wherein W linear is the weight of the linear layer; b linear is the threshold value of the linear layer.

[0092] 2.2) Constructing the neural network error back propagation framework

[0093] The mean square error (MSE) is used as the loss function of the neural network back propagation, and the optimizer is the Adam optimizer.

[0094] 2.3) Setting the basic information of the neural network framework

[0095] 2.3.1) Selecting the levitation gap and current of the electromagnet at time t and a period of time before time t as the input value of the neural network, and the levitation gap at time t+1 as the expected output value Y * Train ,

[0096] 2.3.2) Setting the neural network parameters

[0097] The neural network parameters include the number of hidden layers, the number of hidden layer cells, the output layer error loss function, the back propagation learning rate η, and the training stop error LOSS end , etc.

[0098] 2.4) Using the data samples obtained in step 1) above to train the neural network constructed in steps 2.1) and 2.2) above to obtain a maglev train dynamic response prediction model:

[0099] 2.4.1) Reading the measured maglev train electromagnet levitation gap and current data in step 1) above as the sample set of the prediction model, and obtaining the input value and the expected output value of the neural network from the sample set according to the method of step 2.3.1) above;

[0100] 2.4.2) Initializing the neural network, and assigning initial random values to the weights and thresholds of each neuron;

[0101] 2.4.3) Calculating the input value according to the neural network forward propagation framework built in step 2.1) above to obtain the predicted value Y Train , Y Train = [y L+1 yL+2 …y n ];Then the calculation error LOSS of the output layer is calculated as shown in the following formula (8):

[0102]

[0103] 2.4.4) Compare the calculation error LOSS obtained in the above step 2.4.3) with the training stop error LOSS in the neural network parameters in the above step 2.3.2): if LOSS < LOSS end , the training is completed, and the trained prediction model f end () is obtained; if not, the error is back-propagated, and the error of each neuron in each layer is calculated backwardly; pre

[0104] 2.4.5) Calculate the increments Aw and Ab of the weights and thresholds of each neuron by using the error of each neuron obtained in the above step 2.4.4), and update the weights w and thresholds b of each neuron as w = w + Aw and b = b + Ab, and the training is completed once;

[0105] wherein:

[0106]

[0107] In the above formula (9), η is called learning rate, and the error back-propagation algorithm (BP neural network) can be used to quickly solve

[0108] 2.4.6) Take the weights w and thresholds b obtained in the above step 2.4.5) as the initial values of the neurons in the next round of training, and repeat the calculation process of the above steps 2.4.3) to 2.4.5) again, and continuously modify the weights and thresholds by training so that the error LOSS of the output layer is less than LOSS end , and finally obtain the trained prediction model f pre ().

[0109] 2.4.7) Use the trained neural network model obtained in the above step 2.4.6) to construct a prediction model, and the expression is as follows:

[0110]

[0111] wherein in the above formula (10), f is the predicted value of the suspension gap calculated at time t; f pre () is the trained neural network prediction model; X L (t) is a known measured electromagnet data sequence at time t, and L is the length of X L ​(t) the number of rows including historical information; x(t) is the real data measured at time t, wherein z(t) is the measured levitation gap of the electromagnet at time t; and i(t) is the measured current of the electromagnet at time t.

[0112] The function of the prediction model is to read the levitation gap and current of the electromagnet at previous time points at time t, and combine the measured data at this time to quickly predict the levitation gap of the electromagnet of the maglev train at the next time point.

[0113] 3) The suspension controller reads the gap signal and coil current of the electromagnet, and uses the prediction model to predict the future response of the train, and calculates the feedforward control amount according to the prediction model;

[0114] 3.1) The suspension controller reads the levitation gap and coil current of the electromagnet and reads the historical levitation gap and current, and then estimates the levitation gap of the electromagnet at the future time point through the prediction model in step 2) above

[0115] 3.2) The predicted value of the levitation gap of the electromagnet of the suspension system is used to calculate the feedforward control amount, as shown in the following formula (11):

[0116]

[0117] wherein K ff is a feedforward gain coefficient; z(t) is the measured levitation gap of the electromagnet; and z0 is a rated levitation gap, which is 10 mm.

[0118] 4) The feedback control amount is calculated by reading the measured levitation gap and acceleration of the sensor through the feedback module of the suspension controller; the PID algorithm is used for feedback control, and the proportional (P), integral (I) and differential (D) linear combinations of the deviation between the measured value and the target value of the levitation gap of the electromagnet are combined into a control amount; the integral term of the acceleration of the electromagnet is used to replace the differential term of the levitation gap, and the expression of the feedback control logic of the PID algorithm is:

[0119]

[0120] wherein K p is a proportional coefficient; K i is an integral coefficient; and K d is a differential coefficient, z(t) is the measured levitation gap of the sensor, a(t) is the measured acceleration of the electromagnet, and z0 is a rated levitation gap.

[0121] ​5) The suspension controller outputs the control signal after superimposing the feedforward control amount and the feedback control amount. That is, the feedforward control amount in step 3) above and the feedback control amount in step 4) above are superimposed to obtain the control amount u(t+1) of the predictive control, as shown in the following equation (13):

[0122] u(t+1) = η ff u ff (t+1) + η fb u fb (t+1) (13);

[0123] wherein u ff (t) and η ff in the above equation (13) are the feedforward control amount and the weight coefficient thereof, and u fb (t) and η fb are the feedback control amount and the weight coefficient thereof.

[0124] The present application has a reasonable concept, and a prediction model is established based on the dynamic response of the maglev vehicle under the aerodynamic load, so that the coupling effect between the components of the train and the influence of external disturbances are considered, and the strong nonlinear and coupling characteristics of the maglev train are better reflected. Furthermore, the algorithm combines the advantages of feedforward control and feedback control, can effectively suppress the fluctuation of the electromagnet of the maglev train under the aerodynamic load, and makes the train run more stably.

[0125] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A neural network predictive control method for a high-speed maglev train suspension system, characterized in that, Comprise the following steps: 1) According to the levitation sensor deployed on the upper part of the maglev train, the dynamic response of the maglev train in the running process is obtained; the coil current of the electromagnet in the chopper is read by the levitation controller on the train; at the same time, these data are recorded as sample set for constructing prediction model; 2) Use the recorded samples to train the neural network and construct the maglev train response prediction model to realize the online accurate prediction of the future response of the train; Specifically, the following steps are included: 2.1) Construct the neural network forward transmission framework, specifically, set the improved recurrent neural network-long short term memory neural network to construct the neural network forward transmission prediction framework; 2.2) Construct the neural network error back propagation framework; 2.3) Set the basic information of the neural network framework; The specific steps are as follows: 2.3.1) Selecting the solenoid suspension gap and coil current at time t and a period of time before time t as the input values of the neural network, and the suspension gap at time t+1 as the desired output value , ; 2.3.2) setting neural network parameters; and the neural network parameters include the number of hidden layers, the number of hidden layer cells, the output layer error loss function, the back propagation learning rate η and the training stop error LOSS in the neural network end ; 2.4) Use the data samples obtained in step 1) to train the neural network constructed in steps 2.1) and 2.2), and obtain the maglev train dynamic response prediction model; 3) The levitation controller reads the electromagnet suspension gap and coil current signal, and uses the prediction model to predict the future response of the train, and calculates the feedforward control amount accordingly; 4) The levitation controller reads the electromagnet suspension gap and acceleration signal, and calculates the feedback control amount accordingly; 5) The levitation controller outputs the control signal after superimposing the feedforward and feedback control amounts.

2. The neural network predictive control method for the suspension system of high-speed maglev trains according to claim 1, characterized in that, The step 1) in the step of obtaining the dynamic response data is: 1.1) Determine the data required for constructing the maglev train response prediction model Select the suspension gap and coil current data of the electromagnet of the maglev train as the sample set of the prediction model; That is, arrange the levitation sensor on each electromagnet of the maglev train to measure the electromagnet suspension gap when the train is running, and use the levitation controller to read the coil current data of the electromagnet in the chopper; 1.2) Collect the data samples required for constructing the prediction model That is, select the running conditions of the train, and record the time series of the suspension gap and coil current of the electromagnet of the maglev train under various motion conditions.

3. The neural network predictive control method for the suspension system of high-speed maglev trains according to claim 1, characterized in that, The specific steps of step 2.4) are as follows: 2.4.1) Read the measured maglev train electromagnet suspension gap and current data in step 1) as the sample set of the prediction model, and obtain the input value and expected output value of the neural network from the sample set according to the method of step 2.3.1); 2.4.2) Initialize the neural network, and give initial random values to the weight and threshold of each neuron; 2.4.3) Calculate the input value by the neural network forward propagation framework built in step 2.1) to obtain the electromagnet suspension gap prediction value Y Train , Then, the calculation error LOSS of the output layer is calculated as shown in the following formula (8): （8）； 2.4.4) compare the calculated error LOSS from step 2.4.3) with the training stop error LOSS end end pre then the training is finished and the trained prediction model f is obtained; if not, the error is backpropagated and the error of each neuron of each layer is calculated backwards​​ 2.4.5) Calculate the weight increment Aw and threshold increment Ab of each neuron by the error of each neuron obtained in step 2.4.4), and update the weight w and threshold b of each neuron to w=w+Aw, b=b+Ab, and once the training is completed; 2.4.6) Take the weights w and thresholds b trained in step 2.4.5) as the initial values of the neurons in the new round of training, and repeat the calculation process of steps 2.4.3) to 2.4.5) again, and constantly modify the weights and thresholds by training so that the output layer error LOSS < LOSS end , and finally obtain the trained prediction model f pre (). 2.4.7) Use the trained neural network model obtained in step 2.4.6) to construct the prediction model, which is expressed as: （10）； wherein the above formula (10) is calculated at time t is the predicted value of the suspension gap calculated at time t f pre () is a trained neural network prediction model; X L (t) is a known measured electromagnet data sequence at time t, L is the number of historical information included in X L (t); x(t) is the real data measured at time t, where z(t) is the measured electromagnet suspension gap at time t, and i(t) is the measured electromagnet current at time t.

4. The neural network predictive control method for the high-speed maglev train suspension system according to claim 1, wherein, The specific steps of step 3) are as follows: 3.1) The levitation controller reads the electromagnet suspension gap signal and coil current signal as the input value of the prediction model, and estimates the future response of the suspension system through the prediction model in step 2); 3.2) Calculate the feedforward control amount using the predicted value of the suspension system, as shown in the following formula (11): （11）； wherein K ff is a feedforward gain coefficient; is a predicted suspension gap; z0is a rated suspension gap, taken as 10 mm; the calculation of the feedforward control amount based on the predicted value of the suspension system is not limited to the form shown in equation (11) and can be replaced by other forms as needed.

5. The neural network predictive control method for the high-speed maglev train suspension system according to claim 1, wherein, The specific steps of the step 4) are as follows: The feedback module of the suspension controller reads the suspension gap and acceleration measured by the sensor to calculate the feedback control amount; the PID algorithm is used for feedback control, the proportional (P), integral (I) and differential (D) linear combinations of the deviation between the electromagnet suspension gap measurement value and the target value are combined into the control amount; the integral term of the electromagnet acceleration is used to replace the differential term of the suspension gap, and the expression of the feedback control logic of the PID algorithm is: （12）； where K p is a proportional coefficient; K i is an integral coefficient; K d is a derivative coefficient, z(t) is the sensor measurement of the levitation gap; a(t) is the sensor measurement of the magnet acceleration; and z0 is the nominal levitation gap.

6. The neural network predictive control method for the high-speed maglev train suspension system according to claim 1, wherein, The specific steps of the step 5) are as follows: The feedforward control amount in the step 3) and the feedback control amount in the step 4) are superimposed to obtain the control amount of the predictive control, as shown in the following formula (13): （13）； wherein u ff (t) and η ff are feedforward control quantities and their weight coefficients, u fb (t) and η fb are feedback control quantities and their weight coefficients.

Images